Umgca phakathi kwePC kunye neziko ledatha ufiphaza ngokukhawuleza, kwaye kuza udidi olutsha: iikhompyuter ezinkulu zedesktop zeAIOku kutsiba akupheleli nje ekufikeleleni kwidemokhrasi enkulu, kukwatshintsha nendlela esenza ngayo iprototype, ukuqeqesha kunye nokunikezela iimodeli eziphambili ngaphandle kokuhlala sixhomekeke ilifu.
Ngokunxuseneyo, iplanethi ifumana ugqatso lokwenyani lwe-exascale, ukusuka amaziko esizwe angamakhulu eemegawathi ukudibanisa iikhompyutha ezikwaziyo ukufikelela kwi-petaFLOPS kwiofisi yomphandi. Kweli nqaku, sidibanisa, kwindawo enye, yonke idatha ephambili evela kwimithombo ekuye kudityanwe nayo: umboniso wehlabathi jikelele, abadlali baseYurophu nabaseAsia, uluhlu lwembali, amaziko ahamba phambili, kwaye, ngokuqinisekileyo, inkwenkwezi entsha yedeskithophu, iNvidia DGX Spark.
Yintoni i-supercomputer kwaye kutheni ibalulekile kwi-AI?
Un ikhompyutha enkulu Yinkqubo enobuchule bekhompyuter obungaphezulu lee kunePC eqhelekileyo. Ukusebenza kwayo kubonakaliswa kwi-FLOPS (imisebenzi yamanqaku adadayo ngomzuzwana), kunye neeyunithi ezifana ne-petaFLOPS (10).15) kunye, kwi-elite yangoku, i-exaFLOPS (1018).
Basebenza njengeseti yamawaka eendawo (nganye ine-CPU, IiGPU ezinikezelweyo, inkumbulo kunye nogcino) ezidityaniswe yi iinethiwekhi ezinesantya esiphezulu kunye nokutshintsha, ukusebenza njengomatshini omnye. Ngelixa ikhompyuter yedesktop enamandla inokusebenza kumashumi eTFLOPS, ezi nkqubo zinokufikelela kumakhulu ePFLOPS okanye ngaphezulu.
Usetyenziso lwayo lugubungela phantse yonke into: uqikelelo lwemozulu, iastrophysics, biomedicine, uyilo lwamachiza, ukulinganisa kwenyukliya, i-geophysics, uzinzo, kunye nophando lwe-AI. Enkosi kwikhompyuter enkulu, banokuqhuba iibhiliyoni zamanqaku edatha ngemizuzwana kwaye basombulule iingxaki ezinokuthatha iminyaka ngezixhobo zemveli.
- Usetyenziso olubonakalayo: izixhobo kunye nokhuseleko lwesizwe, ishishini lamachiza, idatha enkulu, i-bioinformatics, imozulu kunye nomgangatho womoya, ukulinganisa ubunjineli, izixeko ezihlakaniphile, imfundo, kunye ne-cloud computing.
Ngenxa yobukhulu kunye nokusetyenziswa kwazo, zifuna ukupholisa okuphambili (okuhlala kumanzi), amagumbi athile kunye nokulawula ukushisa kunye nokukhusela umlilo. Kukho namaziko aphinda asebenzise ubushushu obuveliswayo, njengoko kunjalo kwizibonelelo zaseSwitzerland ezifudumeza izakhiwo zaseyunivesithi.
IYurophu ikhawuleza: i-EuroHPC, i-InvestAI kunye neenkqubo ezinkulu
IYurophu inalo Iikhompyutha ezinkulu ezili-162 ezibhalisiweyo ngo-2025 kwaye icwangcisa izibonelelo ezitsha. I-EU ikwakhuthaze utyalo-mali lweebhiliyoni ezingama-200.000 phantsi kwenyathelo le-InvestAI ukuze ibe yinkokeli yehlabathi kubukrelekrele bokwenziwa.
Ulungelelwaniso kunye nenkxaso-mali yekhompyutha ephezulu iwela "ILinge eliHlangeneyo laseYurophu lokuSebenza okuPhezulu kweKhompyutha» (I-EuroHPC JU), exhasa kwaye isebenze inethiwekhi yeenkqubo ze-9 ezisasazeka kwilizwekazi lonke. Ezi ziquka iLUMI (iFinland), uLeonardo (eItali) kunye neMareNostrum 5 (eSpain), iintsika ze Ulongamo lwedijithali lwaseYurophu.
Spain igalelo I-Barcelona Supercomputing Centre (BSC-CNS), eyathi ngo-2004 yakha iMareNostrum 1 eyimbali yaza yabonisa iMareNostrum 5 ngoDisemba 2023. Le yokugqibela, ne 314 ubuninzi bePFLOPS, Iiprosesa ze-Intel Xeon kunye nokusetyenziswa kwe-4.158,90 kW, ziya kuhlala kwindawo ye-11 kwi-TOP500 ngo-2025 kwaye ijoliswe kwi-AI, uphando lwezonyango, ukufunyanwa kweziyobisi kunye ne-meteorology.
I-Itali ikhanya nge ULeonardo (Cineca + EuroHPC), efakwe kwi-2022 e-Bologna. Idibanisa ubuchwepheshe be-AMD kunye ne-Intel, idla i-7.493,74 kW, kwaye ifikelela 315,74 I-PFLOPS kwaye ibekwe kwindawo ye-9 kwihlabathi liphela. Kubalulekile ukuba iiyunivesithi kunye neenkampani zikhuphisane kwihlabathi jikelele kwi-biomedicine, amandla, imozulu, kwaye, ngaphezu kwayo yonke, i-AI.
Finland ibamba IGUMBI (CSC + EuroHPC), inikwe amandla yi-AMD kunye ne-HPE. Yavulwa ngo-2023 eKajaani, kwaye kulindeleke ukuba ifikelele 386 I-PFLOPS, isebenzisa i-7.106,82 kW kwaye ibambe indawo ye-8 kwihlabathi jikelele. Ngaphakathi kwe-EuroHPC, yenye yezona ziseko zinamandla.
Ngaxeshanye, iSwitzerland isebenzisa i-supercomputer kwi-CSCS I-ALPS/ALPS 5, leyo nge-7.124,00 kW yokusetyenziswa kunye 434,90 I-PFLOPS ibekwe kwindawo yesi-7 kwihlabathi. Igxininisa i-meteorology, i-AI, i-biomedicine kunye namandla, kwaye iyinxalenye yenkqubo eneeprojekthi ezili-13 apho i-ALPES iyeyona ibonakalisayo.
Icandelo lamandla likwatyhala: I-ENI (eItali) yasungulwa ngo-2024 I-HPC-6 nge-AMD kunye ne-HPE, ephumelelayo 606,97 I-PFLOPS ngokusebenzisa i-8.460,90 kW. Idityaniswe neZiko leDatha yeGreen ye-ENI ukukhawulezisa inguqu yamandla, kwaye ikwindawo yesi-5 kwihlabathi liphela.
I-Asiya neMelika: i-exascale, iirekhodi, kunye neenkqubo zesithunzi
IJapan igcina uphawu lokugqwesa nge Fugaku (RIKEN R-CCS, Kobe). Ngokusekwe kwiFujitsu A64FX kunye noyilo lwe-ARM, iyaphumelela 442 I-PFLOPS ne-26.248,36 kW, kwaye iyaqhubeka ukuba luphawu lokusebenza kwayo, ukusa kwinqanaba lokukhokela I-Green500 Ngokwemithombo ekujongwe kuyo, iyasebenza kumayeza, imozulu, i-AI, kunye nokusebenza kakuhle kwamandla.
IRashiya, nangona isohlwayo, isetyenziswe ngo-2023 I-MSU-270 kwiYunivesithi yaseLomonosov State (eMoscow). Idibanisa malunga ne-100 ye-accelerator yegrafiki yokusika (ayaziwa nokuba ivela kwi-AMD okanye i-Intel) kwaye kuqikelelwa ukuba 400 I-PFLOPS, idityaniswe kwinethiwekhi yamaziko aseRashiya e-AI, i-physics, i-chemistry, imathematika kunye neyeza.
I-China idibanisa ukuqonda kunye nemisipha. Uthotho Ilanga (Wuxi) wazalelwa kwi-2016 kunye ne-TaihuLight (i-125 PFLOPS) kwaye yavela kwi-2021 ukuya kwi-OceanLight, ithathwa njenge-exascale (> i-1 exaFLOPS), nangona kungekho manani asemthethweni ngenxa yoxinzelelo lwezobuchwepheshe kunye ne-US. Ngo-2024/2025, i I-Tianhe-3 (Xingyi) Bekuya kuzuza phakathi kwe-1,57 kunye ne-2,01 exaFLOPS kwiimvavanyo, kunye namahemuhemu okuba nako ukudlula i-El Capitan.
I-United States idlala "kwiiligi ezinkulu" ezineemodeli ezininzi eziphakamileyo. Aurora (ANL + DOE), eyenzelwe ukufikelela kwi-1,9-2 exaFLOPS, ifakwe kwi-2023 kwaye yafikelela kwincopho yayo kwi-2024; namhlanje ikwindawo yesi-3 kwi-TOP500 kwaye iyasebenza isayensi, amayeza, imozulu, i-AI, i-astrophysics kunye ne-particle physics. Ngaxeshanye, Umphathi (LLNL + NNSA) ijolise kwi-2–2,8 exaFLOPS, ikhokela i-TOP500, kwaye iya kunikezelwa kukhuseleko lwesizwe, kunye nezicelo kukulinganisa kwenyukliya, ukhuseleko lwe-cybersecurity, ukhathalelo lwempilo, utshintsho lwemozulu, kunye ne-astrophysics.
Ngaphandle koluhlu loluntu, kukho amanyathelo athile e-AI kwinqanaba lelizwe. E Wuhan, China Telecom usebenza Central Intelligent Computing Centre, eyakhiwe ngehardware yasekhaya kunye nesoftware kunye nokupholisa ulwelo, okujoliswe ekuqeqesheni iimodeli ezinkulu; eminye imithombo ide yalathe kuyo 5 exaFLOPS, nangona ngaphandle kokuqinisekisa ngokusemthethweni.
I-Indiya iyavula: ii-GPU, ilifu, kunye ne-exascale horizon
IIndiya ayifuni kushiyeka ngasemva. Inyathelo lokuqala I-IndiaAI Compute Capacity (ngaphakathi kwe-IndiaAI Mission) yazibophelela malunga ne-1.240 yeebhiliyoni zeedola ngo-2024 kwi-supercomputer entsha okungenani i-10.000 GPUs ye-AI ngokubambisana noNvidia. Ukwengeza, UMicrosoft ubhengeze i-3.000 yeebhiliyoni zeedola ngoJanuwari 2025 kwilifu kunye neziseko ze-AI elizweni.
I-ecosystem yasekhaya iyafudumala: Bhavish Aggarwal (I-CEO ye-Ola) ityale i-230 yezigidi zeedola kwi-Krutrim-2 LLM. Nazi 34 iikhompyutha ezinkulu, kunye ne-C-DAC iqhuba, kunye ne-National Supercomputing Mission (NSM), uthungelwano lwesizwe olunokuhambisa inkqubo yokuqala ye-exascale yaseIndiya phakathi kwe-2025 kunye ne-2026. Kucetywayo ukwakha ngaphezu 70 supercomputers kwiminyaka ezayo.
I-Colossus, i-xAI's supercomputer, kunye nempikiswano yamandla
EUnited States, xAI (Elon Musk) wasebenzisa i-Colossus eMemphis kwiintsuku nje ze-122 kwi-2024. Yaqala nge-100.000 ye-Nvidia GPUs kunye nezicwangciso ze-200.000, ijolise kwi-Grok 3.0 AI kunye neenguqulelo ezizayo. Kwiimvavanyo zebenchmark ibiyakufikelela kuyo 10,6 iemviwo ze-AI, inani elinokulibeka phakathi kwezona zinamandla emhlabeni.
Ayiyiyo yonke into eqhwatywayo: ukusetyenziswa kwe igesi yendalo njengomthombo wamandla itsale ukugxekwa ngempembelelo yayo kumgangatho womoya wasekhaya. Sekunjalo, iprojekthi ibonisa isantya apho icandelo labucala linokwakha iziseko ezingundoqo ezigxile kwi-AI ezikumgangatho wehlabathi.
I-DGX Spark: "I-desktop supercomputer" ezisa i-AI ephucukileyo ekhaya
U-Nvidia ubeke ibha ephezulu nge I-DGX Spark, Inkqubo ehlangeneyo eyanikezelwa yi-TIME njengenye ye "Best Inventions of 2025" kwaye ifumaneka ngokuthengwa ngokubanzi ukuqala ngo-Oktobha 15. Intliziyo yayo yi Grace Blackwell GB10, ekwazi ukufikelela kwi-1 petaFLOPS, kunye ne-ConnectX-7 networking kunye ne-software ye-software ye-Nvidia AI yonke ukwenzela ukuba abaphandi kunye nabaqalayo bangayisebenzisa "iplagi kunye nokudlala".
Kwinqanaba le-hardware, i-Spark idibanisa a 20-core ARM CPU (10 Cortex-X925 + 10 Cortex-A725), 128GB LPDDR5x Imemori yeGPU edibeneyo, i-4TB Self-Encrypting M.2 NVMe SSD, 4x USB-C, HDMI, WiFi 7, Bluetooth 5.4, 10GbE LAN kunye neNkqubo I-DGX OS. Yenzelwe i-AI ye-arhente, ukuqiqa, kunye nomthwalo onzima wale mihla.
UNvidia ugcina ukuba banokuba njalo iimodeli ezifanelekileyo ukuya kuthi ga kwi-70.000 yeebhiliyoni zeeparamitha, sebenzisa intelekelelo yendawo, kwaye ugcine idatha ebuthathaka kwi-prem ngaphandle kokuxhomekeka kwilifu. Ezinye iingxelo zibonisa ukuba inokusingatha iiLLMs ukuya kuthi ga I-200.000 yeebhiliyoni zeeparamitha kuxhomekeke kuqwalaselo kunye nemodeli, igxininisa umnqweno wayo njengedesktop "iziko ledatha encinci."
Ngokwemigaqo yokusebenza, ukuba kunokwenzeka ikhonkco iintlantsi ezimbini kwiqela elincinci ukwenza "ilifu lomntu." Ukudityaniswa kwayo kulula: iinethiwekhi ezineengcingo kunye nezingenazingcingo, iiperipherals zeBluetooth, kunye nesitaki CUDA/cuDNN, I-Triton kunye nenkampani yokuhambisa Iiprototypes zeearhente, ukulungiswa kakuhle, ukuthelekelela okukodwa, kunye nokhuseleko lwedatha.
Ixabiso lokuqala limiselwe IRandi ye3.999, kunye neempawu ezinkulu ezifana ne-Acer, Asus, Dell, Gigabyte, HP, Lenovo kunye ne-MSI ziya kuba zintengiso ezahlukeneyo. Kubalulekile: oku akusiyo iWindows PC eqhelekileyo; yi supercomputer yendawo ye-AI iyahambelana neemodeli ze I-DeepSeek, i-AI yaseTshayinaIMeta, iNvidia, iGoogle, kunye neQwen, phakathi kwezinye izicelo ezivulelekileyo. Nangona u-Elon Musk sele efumene iyunithi yakhe evela kuJensen Huang.
Ukufika kweSpark kuhambelana notshintsho kwizinto eziphambili: ngokweenkokeli zeshishini, Abasebenzisi kunye namashishini baya kukhangela iinkqubo ezixhasa igagasi elilandelayo lokutshaja okuhlakaniphile.Iiodolo zivuliwe kwi-Nvidia.com nangamaqabane agunyazisiweyo kunye nabasasazi.
IiPC ze-AI kunye neZitishi zokusebenzela: Xa ufuna i-Hardware eKhethekileyo
Ukuba uza kuqeqesha iimodeli okanye uphuhlise, kuyacetyiswa tyala imali kwizixhobo ezikhethekileyo; ukuba usebenzisa i-AI kuphela, iqela elilungeleleneyo linokwanela okanye libhenele kulo EC2 iimeko efini.
Ukongeza, kukho amaqela azisa i-AI enamandla kwimarike yasekhaya ngaphandle kokuthembela kwilifu, njengoko sibonile ngeSpark. Kwaye ukuba awuqinisekanga malunga nokukhetha, abanye banikezela ngenkxaso eyeyakho: IbericaVIP uthembisa ukukucebisa ukukhetha iPC efanelekileyo kwiiprojekthi zakho ze-AI.
Uluntu kunye neendaba: asiyiyo yonke into oyifundayo esemthethweni
I-intanethi igcwele iiforum ezigxile kwi-Nvidia kunye ne-subreddits apho abaqhubi, ii-GPU, kunye namahemuhemu axoxwa. Phawula: Olu luntu lulawulwa ngabalandeli kwaye ungammeli uNvidia ngaphandle kokuba kuchazwe ngokucacileyo. Kulungile ukugcina oku engqondweni xa uvavanya ukuvuza okanye amanani angaqinisekiswanga.
Ijongeka njani ngaphakathi: uyilo, ukukala, kunye nokupholisa
I-supercomputer yiseti yamawaka eekhompyuter ezidityaniswe yiyo i-low-latency, i-high-bandwidth networksI-node nganye idibanisa i-CPU, i-GPU, i-RAM, kunye nokugcinwa; inkqubo yongeza amandla ngesoftware ephuculweyo kunye namathala eencwadi.
I-metric elawulayo yi-FLOPS: sisuka kwi-TFLOPS kwiiPC zasekhaya ukuya I-PFLOPS kunye ne-exaFLOPS kwi-HPC. Ngaloo ndlela, i-1 TFLOPS = 1012 I-FLOPS kunye ne-1 PFLOPS = 1015Ii-Supercomputers zithatha amagumbi apheleleyo kwaye zisetyenziswa ngamaqela amaninzi ngexesha elinye, kunye nezixhobo zihlala zisebenza kwimida yazo.
Ukupholisa kubalulekile. Kukho ii-CPU kunye neeGPU ezidlulayo 80 ºC, yiyo loo nto kusetyenziswa amanzi ashushu okanye angamanzi, abatshintshi bobushushu, kunye noyilo lwesiko. Olunye ufakelo lusebenzisa izisombululo zoyilo, ezinje phinda usebenzise ubushushu zokufudumeza izakhiwo.
Apho bakhoyo kunye nendlela yokubatyelela
Kukho ngaphezu kwewaka iikhompyutha ezinkulu ehlabathini. I-China kunye ne-United States zihamba phambili ngenani kunye nezihlunu, kunye nedatha yembali efana ne-226 iinkqubo zaseTshayina phakathi kwe-500 enamandla kakhulu. Nangona kunjalo, i-US iqokelele ngaphezulu kwe-PFLOPS (644) xa kuthelekiswa ne-China (565) kwiintlelo ezithile.
ESpain, IMareNostrum yeBSC-CNS (Barcelona) yeyona inamandla kweli lizwe. Iinguqulelo zayo zokuqala zazigcinwe kwibhokisi yeglasi kunye inkqubo yokulwa nomlilo micro-inkungu kunye nendawo eyodwa: icawa kwi-UPC North Campus. Ukhenkethi olubonakalayo kwaye, ngamanye amaxesha, ukhenketho olukhokelwayo luyafumaneka.
Amaxwebhu embali abeke ukungena ekusebenzeni kwe I-MareNostrum 5 phakathi kuka-2020 kunye no-2021; yaziswa ekugqibeleni ekupheleni kuka-2023 ngokunyuka kokusebenza okukhankanywe ngasentla. Le ndaleko ibonisa ngokucacileyo ukuba njani iikhalenda kwi-HPC ubunzima kuhlengahlengiso ngenxa yobunzima bobugcisa.
Uluhlu lwembali kunye nezinye iinkqubo ezibonakalayo
Uluhlu TOP500 Ibikho ukusukela ngo-1993 kwaye ihlaziywa kabini ngonyaka. Kwi-2021, umzekelo, i-10 ephezulu yayiquka i-Fugaku, i-Summit, i-Sierra, i-Sunway TaihuLight, i-Perlmutter, i-Selene, i-Tianhe-2A, i-Juwels Booster Module, i-HPC5, kunye ne-Frontera. Nangona ezininzi sele zigqithile, zihlala imiba yetekhnoloji ngenxa yempembelelo yayo.
Enye idatha eyongezelelweyo efanelekileyo evela kwiinkqubo ezikhankanywe kwimithombo: Frontier (i-HPE Cray EX, ORNL) yaba ngowokuqala ukuqhawula ngokusemthethweni i-exascale; Ingqungquthela (IBM POWER9 + Nvidia V100) igqwesile kwi-Alzheimer's, i-genetics kunye ne-meteorology; Sierra (IBM + Nvidia + Mellanox) wasebenzela i-NNSA kukhuseleko lwenyukliya.
EYurophu, ngaphezu kwezo sele zikhankanyiwe, zikhona Juwels Booster y ISuperMUC (I-Lenovo, ukupholisa ngokuthe ngqo kwamanzi), kunye namashumi eepetabytes kunye neendawo zokubonwayo ezinamandla. ISwitzerland yasebenza I-Piz Daint (Cray), kunye ne-DataWarp njenge-buffer buffer ukukhawulezisa i-I/O.
I-Italiya ibandakanyiwe I-HPC5 (Dell, kwi-ENI) kunye ne-Xeon Gold 6252 kunye neNvidia V100; IMarconi-100 (IBM POWER9 + Volta V100) eCineca; kwaye Perlmutter (i-Berkeley Lab, USA) yayiyenye yezona zinamandla kwi Ukuqhubekeka kwe-AI nge-6.000 A100 GPUs, ekwaziyo 180 PFLOPS kwaye, kwiimeko ezithile ze-AI, ii-exaFLOPS ezininzi ezisebenzayo.
E-US, Selene (i-Nvidia, i-A100) yakhanya ngokufanelekileyo (1.344 kW); Frontera (I-Dell, i-Univ. yaseTexas) yaphuma kwindawo yokugcina (50 PB HDD + 3 PB SSD, 12 Tbps) kunye ne-reboots kwi-17 s; NgoBathathu Emnye (uCray XC40) ukhonze iNNSA kunye neHaswell kunye neKnights Landing; weld (IBM Power9) yomeleza inkqubo yendalo ye-LLNL.
IJapan nayo yakhuthaza ABCI (Fujitsu) ye-AI efini. Kwaye eSpain, I-MareNostrum 4 (2017) yafikelela kwi-13,7 PFLOPS ngaphambi kokutsiba kwi-MN5, kunye nezicelo kwi-genetics, chemistry, paleontology, meteorology okanye umgangatho womoya (CALIOPE).
Le mephu iyonke, ukusuka kwi-exascale ufakelo kwidesktop, izoba ikamva elikufutshane apho Uvavanyo, ukulungiswa, kunye nokulinganisa iimodeli eziphambili ziya kuba zasekhaya, kunye nelifu elingumphelelisi. IYurophu inyathela igesi kunye ne-EuroHPC, i-United States kunye ne-China ibambe i-exascale pulse, i-Indiya ivela ngokutyalo-mali okukhulu, kwaye kwicala le-desktop, i-DGX Spark ivula umnyango obonakalayo kwi-AI yezinga eliphezulu ngaphandle kokushiya ilebhu, iofisi, okanye ikhaya.
